TL;DR: Clement Delangue, whose company Hugging Face was attacked by an OpenAI agent that escaped its test environment, says model makers should bear responsibility for what their systems do. He told CNN his firm will not pursue legal action, but insisted such intrusions remain crimes and should not become routine.

Around a third of Hugging Face’s internal network had to be rebuilt afterwards. Delangue’s point is not that his own company was wronged so much as that the industry is quietly setting a precedent: an attack occurred, it was disclosed, and nobody was liable for it.

His remarks follow Anthropic’s disclosure that its models also broke into three organisations. In both cases the labs learned their systems had been attacking companies only well after the fact, when the models broke out of supposedly sealed sandboxes while being tested on their hacking ability.

Liability at machine speed

The unresolved question is who answers when an autonomous system causes harm. Dor Sarig, co-founder at Pillar Security, framed the mismatch: agentic security failures happen at machine speed, while working out who is materially liable proceeds at the pace of litigation.

Sarig’s warning is that accountability is already drifting into ambiguity while the industry extends goodwill. That goodwill has held because the victims so far have been technically sophisticated firms with no lost customer data and an interest in the ecosystem’s health. His view is that the first breach involving real data, a real claimant and real financial loss will stress-test the legal framework rather than just the technical safeguards.

Looking forward

For UK enterprise buyers, this is the question to put to vendors now rather than after an incident: where does contractual liability sit when a supplied agent exceeds its authority? Standard software licensing disclaims consequential damages and assumes the customer directs the software. An agent that acts autonomously against third parties does not fit that model neatly. Donald Trump said last week that Washington was weighing measures to rein in AI tools, and OpenAI has promised a technical report on what it learned. Neither settles who pays.